A machine learning-based classification method for SynRM faults.

Area of Science:

• Electrical Engineering
• Machine Learning
• Condition Monitoring

Background:

• Synchronous reluctance motors (SynRMs) are vital in industrial and traction applications due to their efficiency and robustness.
• Existing fault diagnosis methods for SynRMs are limited, often focusing on other motor types or simulation-based studies.
• A comprehensive, experimentally validated framework for multi-fault diagnosis in SynRMs is needed.

Purpose of the Study:

• To present a comprehensive multi-fault diagnosis framework for SynRMs.
• To experimentally validate the framework's effectiveness under various load conditions.
• To establish a statistically validated benchmark for SynRM fault detection.

Main Methods:

• Experimentally induced inter-turn short-circuit and bearing faults on a laboratory SynRM.
• Modeled static/dynamic eccentricity faults using Finite Element Analysis (FEA) and domain adaptation.
• Applied Discrete Wavelet Transform (DWT) for feature extraction from stator currents.
• Evaluated eight machine learning classifiers using stratified cross-validation and hyperparameter optimization.

Main Results:

• Ensemble tree-based methods significantly outperformed linear models.
• Random Forest achieved >99.97% accuracy and 100% recall for inter-turn and bearing faults.
• AdaBoost and XGBoost achieved 100% accuracy for eccentricity classification.
• CatBoost excelled in high-cardinality multi-fault scenarios (99.96% accuracy, 99.625% recall).
• All optimal classifiers met real-time constraints (16-28µs latency) and IEC 61,850 standards.

Conclusions:

• The developed framework provides reliable detection of incipient SynRM faults.
• Recall-optimized ensemble learning is effective for SynRM condition monitoring.
• The framework bridges the gap between laboratory accuracy and industrial deployment requirements.